Generates a ranked addressbook from a maildir folder. It can be used in MUA's like aerc or mutt by grepping the list.
Why? No need to manually edit an address book, yet the cached ranking is available extremely fast.
- scans all your emails
- ranks based on both recency and frequency of addresses
- collects from To, Cc, Bcc and From fields
- ranks addresses explicitly emailed by you higher
- configurable output via go templates
- uses the most frequent non-empty display name for each email
- display name can be unicode normalized for search purposes
- filters common "no reply" addresses, additional filters can be added via regexes
- normalizes emails to lower case
- ability to add additional email addresses from a command
- "blazingly fast"*: crunch time for 270k emails is 7s on my machine, grepping from the output is instantaneous
*: compared to original python implementation for crunching (see Behind the scenes below) and compared to using notmuch query for address completion
The easiest way to install is running:
go install github.com/ferdinandyb/maildir-rank-addr@latest
At the minimum, you need to specify where your maildir formatted email are:
maildir-rank-addr --maildir=~/.mail
For most use cases, it likely only needs to be run once or twice a day (cronjob or systemd timer).
Supported flags:
--addr-book-cmd string optional command to query addresses from your addressbook
--addr-book-add-unmatched if cmd is stated, determine wether to add unmatched addresses at the end of the file (true or false)
--addresses strings comma separated list of your email addresses (regex possible)
--config string path to config file
--filters strings comma separated list of regexes to filter
--maildir string path to maildir folder
--outputpath string path to output file
--template string output template
maildir
The path to the folder that will be scanned. No default is set for this.
outputpath
By default results are output to
$HOME/.cache/maildir-rank-addr/addressbook.tsv".
addresses
List of your own email addresses. If you do not provide your own addresses, classification based on your explicit sends will not be possible!
template
Uses go's text/template to configure output for each address (one line per address).
Available keys:
Address
Name
NormalizedName: same as Name, but unicode normalized
Names
Class
FrequencyRank
RecencyRank
TotalRank
ClassCount
ClassDate
Default: {{.Address}}\t{{.Name}}
filters
List of regexes. If an address is matched against a regex, it will be excluded from the output. The regex is matched against the entire email address.
Note that we already filter out addresses, where the local part (the part before the @) matches any of these strings:
"do-not-reply",
"donotreply",
"no-reply",
"bounce",
"noreply",
"no.reply",
"no_reply",
"nevalaszolj",
"nincsvalasz",
addr-book-cmd
Optional command to fetch email addresses and names, the output it returns must have an email address first, followed by a tab space and and the desired name, the name must end in a tab space or a newline for the command to work, this can be useful for integrating with command line addressbooks such as abook or khard
abook --mutt-query "s"
khard email -p --remove-first-line
config
Path to a config file to be loaded instead of the defaults (see below).
Besides the flags, toml formatted configuration file is also possible. It's
first looked for at $HOME/.config/maildir-rank-addr/config and then the
current working directory.
Complete example configuration with the default (aerc compatible) template:
maildir = "~/.mail"
addresses = [
"[email protected]",
"[email protected]"
]
filters = ["@spam.(com|org)"]
outputpath = "~/.mail/addressbook"
template = "{{.Address}}\t{{.Name}}\t{{.NormalizedName}}"
Put something like this in your aerc config (using your favourite grep):
address-book-cmd="ugrep -jP -m 100 --color=never %s /home/[myuser]/.cache/maildir-rank-addr/addressbook.tsv"
(-j is smart case insensitive, and needs to be combined with -P for UTF-8).
Since aerc only uses the first two of the tab separated columns any other
column can be added to help with search or to combine with external tools. For
example adding NormalizedName as the third column will allow you to type
"arpad", and still find and use the entry for "Árpád X" who uses accents in his
name properly, and "Arpad Y" who conformed to ASCII for some reason.
Note that address-book-cmd is not executed in the shell, so you need to hard
code the path without shell expansion.
If you are using aerc with [compose].edit-headers=true you need integrate
with your editor (e.g. with vim), instead of the above.
This is an example using fzf and the fzf.vim plugin. Add it to for example
~/.vim/after/ftplugin/mail.vim to load it only for eml files. The example has
been optimized for aerc with [compose].edit-headers=true. To insert
addresses in the To: field, take your cursor to the line containing To: and
press <leader>a. Use tab to select multiple addresses in the pop-up.
function! InsertAddressAerc()
call fzf#run(fzf#wrap("insertaddress", {
\ 'source':'cat ~/.cache/maildir-rank-addr/addressbook.tsv',
\ 'sink*': function("InsertContactsLine"),
\ 'options': '--no-sort -i --multi'
\}))
endfunction
function! InsertAddress()
call fzf#run(fzf#wrap("insertaddress", {
\ 'source':'cat ~/.cache/maildir-rank-addr/addressbook.tsv',
\ 'sink': function("InsertContact"),
\ 'options': '--no-sort -i'
\}))
endfunction
function! InsertContactsLine(names) abort
for name in a:names
call InsertContactLine(name)
endfor
endfunction
function! InsertContactLine(name) abort
let [address, name; rest] = split(a:name,"\t")
call append(line('.'), ' ' . name . " <" . address . ">,")
endfunction
function! InsertContact(name) abort
let [address, name; rest] = split(a:name,"\t")
exec 'normal! a' . name . " <" . address . ">\<Esc>"
endfunction
nnoremap <leader>a :call InsertAddressAerc()<CR>
nnoremap <leader>A :call InsertAddress()<CR>
Ranking is actually done by first classifying and then ranking within class.
First we go through each email found in your maildir and for each address found in any of the address headers we assign a class, based on whether the sender is you or not and which type of header the address was found in:
- 2: from address is yours, address found in To, or Bcc,
- 1: from address is yours, address found in Cc,
- 0: From fields and anything else.
For each unique address seen, we record a class dates (the date of the latest email in which that address was assigned class X) and class counts (the number of times in which that address was assigned class X). The unique address itself also get assigned a class, which is the highest class it was seen in.
Bob writes Alice 3 letters and Alice answers one of them. When Alice runs the software both email addresses will receive 3 counts to class 0 and have the class date for 0 at the latest email. Alice now answers one of Bob's emails and Cc-s Eve. Alice's address receives another count for class 0 and the date is also updated. Eve's address receive a count for class 1 and Bob's email receives a count for class 2 with the class 2 date being set to this latest email. This puts Bob's email address as class 2 as that is the highest class it has, Eve's at class 1 and Alice's own address in class 0.
The addresses are then ranked by their highest class based on only the highest class's count and date. The output will be structured so class 2 emails are on the top, then class 1 email and class 0 emails are at the bottom. This solves two things: firstly, all mailing-lists and marketing which you get but only read will be available to send to, but are guaranteed to be at the very bottom of search results. Secondly, even if you accidentally replied to a newsletter email which you get daily, it's class count will be just 1, so even though it's a bit higher up now due to the reply, it will still feature at the bottom of it's class and not get conflated by having received hundreds of email from the address pretty recently.
Frequency rank: The emails are ordered according to the class count and their frequency rank becomes their place in this ordered list, with the highest count receiving a rank of 0. In case of equal counts the order is the alphabetical order of the email addresses.
Recency rank: Similar, the emails are ordered again, now based on their class date, where the most recent email receives the rank of 0. In case of equal dates the order is the alphabetical order of the email addresses.
Total rank = Frequency rank + Recency Rank
The output is then generated by printing class 2 address from lowest to highest rank, then class 1 addresses from lowest to highest and finally class 0 addresses from lowest to highest. In case the total ranks are equal the order is the alphabetical order of the email addresses.
The amount of email I have seems to grow approximately linearly and the amount of email addresses also more-or-less, but with a much-much smaller coefficient. Compared to needing to grep the email headers caching the unique address leads to a 250x compression. Since grep retains ordering of results in a file, it also makes sense encoding rankings by simply ordering the addresses.
You can generate these images for yourself using the python script stats/generateEmailStatistics.py.
The stats folder also includes the original PoC implementation of this in
python (stats/generateAddressbookMaildir.py) which takes a whopping 36
minutes to complete the same task, compared to this implementation's 10
seconds.
Please see contribution guidelines.
- maildir2addr: somewhat similar address book generator
- notmuch-addrlookup-c: address lookup from notmuch
- addr-book-combine: for combining generated addressbooks with hand currated ones, like khard
Some functions for parsing email was taken from aerc.